This year, ORX set up a group of operational risk professionals to specifically look at the impact of machine learning on operational risk. Together with our Machine Learning Working Group, we collected examples of use cases of the application of machine learning to operational risk.
Our analysis showed that there are three main areas where it can be used to good effect:
- Fraud detection
- Text-mining for data augmentation
- Data quality assurance
Use 1: Fraud detection
The use of machine learning techniques for fraud and anti-money laundering detection is a well-known, almost “classic” application of machine learning to manage these operational risks.
The labelling of financial transactions as either suspicious or innocuous is one of the prime examples of how advanced analytics has sped up processes and reduced resource requirements.
Machine learning techniques can help improve these processes further and reduce false-positives and false negatives (alerts of fraud or money laundering that turn out to be false alarms or instances where fraudulent transactions are overlooked).
Example use case: Card not presented fraud model
The firm developed a model to detect fraudulent transactions in real-time where the card was not presented at the time of transaction.
This was an internal proof of concept model developed to replace a vendor model.
The motivation for using machine learning techniques was the extremely large data volumes that needed to be handled, which meant even small improvements in model prediction had a high potential for cost savings.
Use 2: Text mining for data augmentation
Effective operational risk management relies on large amounts of data for risk quantification and risk management. This includes internal loss data, control libraries, internal risk indicators and other internal data, as well as external loss data and macroeconomic data.
Maintaining these datasets can be time-intensive, especially where it involves the categorisation of individual entries. We have seen financial firms using machine learning techniques to analyse free-text descriptions, for example of loss events, to complete the categorisation of data entries.
Example use case: NLP tagging of loss incidents
The firm used NLP tagging of losses to infer root causes from already existing tags and free-text descriptions. Within the internal dataset of loss incidents, some attributes or dimensions of the incidents had been added as additional requirements over time and therefore had only been collected from a certain time onwards.
Understanding root causes of incidents enables better risk mitigation strategies, but analysing root causes individually from free text is a time-intensive, manual and somewhat subjective process.
Use 3: Data quality assurance
With growing datasets, the data and quality assurance become more time-intensive. Machine learning techniques can help reduce this effort by identifying duplicated entries and identifying outliers more accurately.
Example use case: Idiosyncratic event identification
The firm used an isolation forest model as part of its capital allocation model. The model was designed to identify events unlikely to occur again because of enhanced governance or controls. The model was still being implemented at the time this report was written.
Making a business case for machine learning in op risk
These use cases are taken from our recent white paper. Read the white paper, Machine learning in operational risk: Making a business case for its practical implementation, to find out more.